Meta-learning instant neural graphics
Thesis type RESEARCH
Description Instant Neural Graphics Primitives (Instant NGP)  have shown great success in solving 3D tasks like view synthesis, surface reconstruction and general signal representation. Despite their speed, they require to solve an optimization problem at test time, which still limits their scalability, and they possess no prior knowledge about the target signals. In this thesis, the candidate will study meta-learning techniques to provide an initialization point with prior knowledge to Instant NGP. This is expected to provide benefits such as accelerated convergence, higher-quality solutions or improve the rate-distortion performance of the representation if used for compression, as suggested by evidence on older implicit representation models [2,3].
 Thomas Muller, Alex Evans, Christoph Schied, Alexander Keller, "Instant Neural Graphics Primitives with a Multiresolution Hash Encoding", ACM Transactions on Graphics 2022
 Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P. Srinivasan, Jonathan T. Barron, and Ren Ng, "Learned initializations for optimizing coordinate-based neural representations", CVPR 2021
 Francesca Pistilli, Diego Valsesia, Giulia Fracastoro, Enrico Magli, "Signal Compression via Neural Implicit Representations", ICASSP 2022
Required skills The candidate is required to have familiarity with neural networks and pytorch.
Deadline 03/06/2023 PROPONI LA TUA CANDIDATURA